Live Notes of Evelina Fedorenko’s Keynote @EMNLP 2021

Nov 8, 2021

Speaker: Evelina Fedorenko (Associate Professor in Neuroscience&NLP at MIT)

Notetaker: Zhijing Jin (PhD at Max Planck Institute & ETH; @ZhijingJin)

Motivation: Marr’s 3 Levels

For artificial intelligence, David Marr proposes that we need to reach three levels: the computational, algorithmic, and implementational (Marr, 1976).

• Computational level: neuroscience to decode human brains [Expertise of the speaker]
• Algorithmic level: AI research to model the representations, reasoning, etc.
• Implementational level: deployment/physical realizations of AI models

2 Key Messages of the Talk

In this talk, there are two key messages discovered from the marriage of neuroscience and NLP:

1. The language system does not support non-linguistic cognitive abilities, but closely linked to social cognition ability.
2. Language is not suitable for complex thoughts, but optimized for efficient communication.

Language vs. Other Cognitive Abilities

There are many cognitive abilities that our brain can achieve. Language is one of the many. So how does language differ from/relate with other cognitive abilities?

Various cognitive abilities of human brains. Copyright by Evelina Fedorenko.

The language system:

• does not support non-linguistic cognitive abilities
• is intimately linked with the system that supports social cognition

To keep in mind:

1. Specialization is not innateness
2. The processing of language does not involve computations different from other abilities in our brain
3. The effect of language to the development of thought is limited

What are properties of language?

1. Language is not suitable for complex thought => the language system has a relatively short temporal receptive window.

• The language system is not sensitive to structure at the discourse level
• The temporal receptive window appears to be between 6–10 words

2. Language does not rely on abstract syntax

• No syntactic “hubs”: syntactic processing is distributed across the language network
• No syntax selectivity: every syntax-responsive cell population or brain area is robustly sensitive to word meanings.

Research of the speaker: They separate syntax and word meanings in text, and test how these texts affect the active brain regions.

As shown in the following figures, there are four settings:

1. Text with both syntax and meanings
2. Text with only syntax but not meanings
3. Text with no syntax but with meanings
4. Text with neither syntax nor meanings

The results are

• syntactic processing is distributed across the language network
• every syntax-responsive cell population or brain area is robustly sensitive to word meanings

Experiment settings by varying the presence of syntax and meanings. Copyright by Evelina Fedorenko.

Activated parts of human brains under the four experiment settings. Copyright by Evelina Fedorenko.

Language can be applied to any type of combinatory algorithms, not just syntax.

What is the computational goal of language?

• Language is very efficient for communication / information transfer
• Language processing is fundamentally predictive

Copyright by Evelina Fedorenko.

Language & Intelligence

“Increased in social intelligence” leads to “Increased need to communicate more complex ideas”, and, reciprocally, improved language ability enhances social intelligence again.

Conclusion

Two key messages of this talk:

1. The language system does not support non-linguistic cognitive abilities, but closely linked to social cognition ability.
2. Language is not suitable for complex thoughts, but optimized for efficient communication.

Copyright by Evelina Fedorenko.

Important future work:

1. Collect quality data
2. Dissect high-performing models
3. Link language models with models of reasoning

Three future directions forward. Copyright by Evelina Fedorenko.